ZerdaZerda

config

On this page

All configuration dataclasses and enumerations for the router system, centralized so the entire system can be tuned from one RouterConfig object.

Enumerations

SimilarityMetric(str, Enum)

COSINE = "cosine", DOT_PRODUCT = "dot_product", EUCLIDEAN = "euclidean". Consumed by SemanticScorer.

ExecutionMode(str, Enum)

SINGLE = "single", SEQUENTIAL = "sequential", PARALLEL = "parallel". Consumed by ExecutionEngine.execute.

AggregationStrategy(str, Enum)

FIRST_WINS = "first_wins", VOTING = "voting", MERGE = "merge". Consumed at HierarchicalRouter construction time to select an aggregator instance — see the Configuration page for the important caveat that VOTING has no dedicated implementation.

ConfidenceLevel(str, Enum)

HIGH = "high", MEDIUM = "medium", LOW = "low", NONE = "none". Produced by ConfidenceEvaluator.evaluate.

LogLevel(str, Enum)

DEBUG = "DEBUG", INFO = "INFO", WARNING = "WARNING", ERROR = "ERROR". Consumed by RouterLogger.

All enums inherit from both str and Enum, so they compare equal to their string values (e.g. ConfidenceLevel.HIGH == "high" is True) and serialize cleanly with json.dumps when accessed via .value.

Sub-config dataclasses

Each is a plain @dataclass with no methods beyond the generated __init__/__repr__/__eq__. All fields have defaults, so every sub-config is constructible with zero arguments.

ClassPurposeFull field reference
EmbeddingConfigEmbedding model settingsConfiguration → EmbeddingConfig
ScoringConfigHybrid scorer weights and confidence thresholdsConfiguration → ScoringConfig
ExecutionConfigExecution mode, retries, aggregationConfiguration → ExecutionConfig
CacheConfigRoute-decision cache behaviorConfiguration → CacheConfig
FeedbackConfigAdaptive feedback loop behaviorConfiguration → FeedbackConfig
ObservabilityConfigLogging, tracing, metricsConfiguration → ObservabilityConfig

RouterConfig

Python · 11 lines
@dataclass
class RouterConfig:
    embedding:      EmbeddingConfig     = field(default_factory=EmbeddingConfig)
    scoring:        ScoringConfig       = field(default_factory=ScoringConfig)
    execution:      ExecutionConfig     = field(default_factory=ExecutionConfig)
    cache:          CacheConfig         = field(default_factory=CacheConfig)
    feedback:       FeedbackConfig      = field(default_factory=FeedbackConfig)
    observability:  ObservabilityConfig = field(default_factory=ObservabilityConfig)
    llm_model:      Optional[str]       = None
    llm_api_key:    Optional[str]       = None
    raise_on_no_match: bool             = False

The single entry point passed to HierarchicalRouter(config=...). See Configuration → RouterConfig for the full field table.

Class methods

RouterConfig.fast() -> RouterConfig

Returns a new RouterConfig() with scoring.top_k = 1, execution.mode = ExecutionMode.SINGLE, cache.enabled = True. All other fields remain at their dataclass defaults.

Time complexity: O(1) — constructs a handful of dataclasses.

RouterConfig.accurate() -> RouterConfig

Returns a new RouterConfig() with scoring.top_k = 3, execution.mode = ExecutionMode.PARALLEL, scoring.llm_weight = 0.20, scoring.semantic_weight = 0.60. Note scoring.keyword_weight is left at its default 0.20, so the three scoring weights after calling accurate() are semantic=0.60, keyword=0.20, llm=0.20 before HybridScorer's internal normalization (they already sum to 1.0 here, so normalization is a no-op in this specific preset).

Examples

1. Basic — defaults:

Python · 2 lines
from zerda.router import RouterConfig
config = RouterConfig()

2. Intermediate — partial override:

Python · 2 lines
from zerda.router import RouterConfig, CacheConfig
config = RouterConfig(cache=CacheConfig(enabled=False))

3. Advanced — preset plus override:

Python · 4 lines
config = RouterConfig.accurate()
config.execution.parallel_timeout = 20.0
config.llm_model = "gpt-4o-mini"
config.llm_api_key = "sk-..."

4. Error handling: RouterConfig performs no field validation (e.g. out-of-range thresholds, negative timeouts are accepted silently) — invalid values will surface later as unexpected scoring/execution behavior rather than a constructor-time error. Validate your own config values if they come from user input.

5. Production-ready — environment-driven config:

Python · 9 lines
import os
from zerda.router import RouterConfig, ScoringConfig, ExecutionConfig, ExecutionMode

config = RouterConfig(
    scoring=ScoringConfig(top_k=int(os.environ.get("ROUTER_TOP_K", "5"))),
    execution=ExecutionConfig(mode=ExecutionMode(os.environ.get("ROUTER_MODE", "single"))),
    llm_model=os.environ.get("ROUTER_LLM_MODEL"),
    llm_api_key=os.environ.get("OPENAI_API_KEY"),
)

Related APIs: HierarchicalRouter, full field reference on the Configuration page.

Related pages